Chat

When you send a message, the agent decides which tools the question needs, calls them, reads what they return and answers. The tool calls are listed above the answer as they run, each with the tool and its input, so you can see where every fact came from.

The model’s reasoning streams into a thinking panel you can open while it works.

A turn has no limit on steps or tool calls: the agent works until the answer is done, you press Stop, or your credits run out. One turn can work for about 14 minutes. If it needs longer, it stops with a note and keeps what it found; reply continue to go on from there.

RBAi chooses the model for each part of the app, so there is no model menu. It uses AI models hosted on Microsoft Azure, and we change them as better ones come out. Each answer’s details name the model and the thinking level it ran at.

Thinking, in the composer, sets how much the model thinks before it answers. You choose it once: it applies to every chat, Galaxy Lab, the co-author, Connect, the site designer and Studio, on every device. A Literature Review has its own level, chosen for each review.

RBAi has built-in tool servers for these sources and calls them live:

Proteins and structures

UniProt, RCSB PDB, AlphaFold DB, InterPro, ESMFold.

Genes and variants

Ensembl, ClinVar, cBioPortal, the UCSC Genome Browser.

Pathways and interactions

Reactome, STRING, WikiPathways, Gene Ontology through QuickGO.

Expression

The Human Protein Atlas, and single-cell datasets and gene expression from CELLxGENE.

Sequences

NCBI BLAST, multiple sequence alignment and phylogenetic trees.

Literature

The RBAi corpus of 11 million papers, PubMed, Europe PMC, OpenAlex, Semantic Scholar and preprints, and the open full text of a paper.

Compute

The Python sandbox, PyMOL and Galaxy.

For anything else, the agent can read the API specification of 64 databases, such as Rfam, ClinGen, ENCODE, CIViC and openFDA, and call the operation it needs. Type @ in the message box to name a database.

The agent searches the web when a question needs current information, such as a recent approval, a guideline or a core facility’s page, and cites the pages it used. Give it an address and it reads that page, PDFs included, for example Read https://www.addgene.org/8453/ and tell me its bacterial resistance.

Type @ in the composer to open the menu. It has these groups:

• Databases and tools: one entry per tool server, such as BLAST, Alignment & trees, AlphaFold DB, UniProt, CELLxGENE, Code sandbox and PyMOL.
• Database APIs: the 64 databases whose API specifications the agent can read.
• Any MCP servers you have added.
• Agent Skills.

Mentioning a tool limits the agent to that tool’s functions for the turn. This is useful when you know the method you want.

A skill is a written procedure for a kind of task. The agent loads a skill’s instructions when a task needs them, or when you pick the skill from the @ menu. RBAi ships with skills that include:

• statistical-analysis, statistical-power and pathway-enrichment
• phylogenetics, biopython, bioservices and genomic-intelligence
• scientific-visualization, matplotlib, seaborn and plotly-interactive-plots
• pymol and galaxy-workflows
• literature-review, peer-review, scientific-critical-thinking and hypothesis-generation
• scientific-writing, research-grants, latex-posters and markdown-mermaid-writing

Your own skills

In Settings → Agent skills, upload a SKILL.md (YAML front matter followed by the instructions) with optional reference files. Your skill then appears with the others.

Attach files to a message, up to 25 MB each. The agent reads what is inside them:

• Structures: .pdb and .ent.
• Sequences: .fasta and .fa.
• Tables and data: .csv, .tsv, .json and .xvg.
• Text, PDF and Word (.docx) files.
• Spreadsheets (.xlsx, .xls, .ods) and slides (.pptx): sheets arrive as tables with their names, slides with their numbers, up to 50,000 characters.
• Images: PNG, JPEG and TIFF. The model sees the image itself; large photos are shrunk to fit.

Old Word (.doc) and PowerPoint (.ppt) files cannot be read. Save them as .docx, PDF or .pptx first.

Ask for an interactive chart, a dashboard or a table, and the agent can write it as an HTML page. The page opens in a panel beside the chat (over it on a tablet or a phone), with Preview and Code tabs, and the chat keeps a card with its title, Open and a download button.

Versions

Every change is saved as a new version with the chat. The switcher reads v2 of 3, and its menu lists each version with the request that made it. On an earlier version, Make v1 the newest saves a copy of it as the newest version, with no model call. A page keeps up to 30 versions and 8 MB, its versions count toward your storage, and deleting the chat deletes its pages.

Edit with AI

In the panel, choose Edit with AI, click an element of the page (or Tab into the page, move with the arrow keys and press Enter), describe the change and press Apply. A change to the whole page needs no selection. While it runs, the panel shows the time and Cancel; then a line such as Saved as v3. 4 lines changed. with Show changes, which opens the code at the changed lines.

If numbers on the changed lines differ and your request did not name them, an amber warning lists the old and new values, with Go back to v2. An edit that fails says why, and the page stays as it was.

You can also ask in the chat, for example In the BLEU chart, add a y-axis title. The agent edits the stored page and saves a new version; it does not write the page again from memory.

What a page can load

A page runs in a sandbox, apart from the app and your account. It may load scripts, stylesheets and fonts from cdn.jsdelivr.net and Google Fonts only, and it cannot send requests, post forms or open frames, so it cannot send the data it shows anywhere. When a page tries, the panel lists the problem, for example Blocked a network request to rest.uniprot.org, and Fix with AI asks for a version that works within these rules.

The panel also shows the page at tablet and phone widths, and has Download (the file is named after the page and its version), Print and Copy. Publish to my page puts one version on your public page. Website

The microphone next to Send records up to 2 minutes. Your words appear in the message box for you to check and edit; nothing is sent until you press Send. English and Hindi are recognised, and the box says which it heard. Recordings are not stored.

RBAi keeps a memory folder for you. After each message it updates profile.md with what lasts: your field and work, and how you want answers (units, statistics, citation style). Every new chat reads it, so you need not repeat yourself; a one-off question changes nothing. Say Remember that I work on zebrafish heart regeneration. or Forget … in any chat, and RBAi changes the files; the answer's work log shows each step. A project has its own folder, shared by its chats, whose facts.md keeps the project's decisions and findings. Your position, department and research interests from Settings → Profile go with every message.

Settings → Memory lists your files: open one to read or edit it, delete it, or use Delete all memory, and the Use memory switch. With memory off, RBAi neither reads nor changes the files, and nothing about you is sent with a chat. The notes RBAi kept before 7 Nov 2026 are in notes.md.

Every chat has its own address, so a reload, Back and Forward open the same conversation, and you can bookmark a chat or open it in a new tab:

• /chatbot/<id> for a chat, and /chatbot/new for a new one;
• /workspace/project/<project>/chats/<id> for a project’s chat;
• /workspace/galaxy_lab/<id> for a Galaxy Lab conversation.

A new chat’s address changes to its own when you send the first message, while the answer streams. A project’s sections have addresses too, such as /workspace/project/<project>/studio. Plain /chatbot opens the last chat you had open in this browser. A chat that was deleted, or that belongs to another account, shows This chat doesn’t exist or isn’t yours, with New chat.

Share, on a result in the chat, opens Send to an RBAi member. Find a member by name, institution or email and press Send: they get a Connect message with the chat’s title and a link to the result as you sent it. Only the two of you can open that link. Send work to a member

In Settings → MCP servers, add an external tool server. Its tools appear in the @ menu, and the agent can call them like the built-in ones.